Nodes/ComfyUI_Wan2_1_lora_trainer/Musubi Cache Latents (Wan)
ComfyUI Node

Musubi Cache Latents (Wan)

A ComfyUI node in musubi-tuner/wan/preprocess with 11 inputs and 3 outputs.

By jaimitoes·Created about a year ago·Updated 12 months ago· 61
Musubi Cache Latents (Wan)
  • trigger_in
  • dataset_toml_path
  • trigger_out
  • status_message
dataset_config_toml
vae_name
vae_dtypebfloat16
devicecuda
batch_size_override0
num_workers_override0
skip_existingfalse
keep_cachetrue
vae_cache_cpufalse
clip_nameNone
Categorymusubi-tuner/wan/preprocess

Inputs (11)

NameTypeDefaultDescription
dataset_config_tomlSTRING
trigger_in*
vae_nameCOMBOVAE Model (Variational AutoEncoder). Select the VAE that is compatible with your base model. The VAE is like the model's "eyes," helping to encode and decode images during the process.
vae_dtypeCOMBObfloat16The data precision type for the VAE (the component that handles images). `bfloat16` is a good option for its balance between speed and quality.
deviceCOMBOcudaChoose the hardware device to use for processing. `cuda` uses your NVIDIA GPU (much faster), while `cpu` uses your computer's main processor (much slower).
batch_size_overrideINT0Override the default batch size used for processing data. `0` means use the default value from the dataset configuration. A higher number processes more items at once, but uses more memory.
num_workers_overrideINT0Override the default number of workers for loading data. `0` means use the default value from the dataset configuration. A higher number can speed up data loading, but consumes more CPU and RAM.
skip_existingBOOLEANfalseIf enabled, the node will skip processing files that already have existing cached versions, saving time and computation.
keep_cacheBOOLEANtrueIf enabled, the processed data will be kept in the cache directory even after this operation completes, so it can be reused later without reprocessing.
vae_cache_cpuBOOLEANfalseIf enabled, the VAE's processed outputs will be cached in your system's main memory (RAM) instead of GPU memory. Useful for saving VRAM, especially with large datasets, but can be slower if you have a slow CPU or hard drive.
clip_nameCOMBONoneCLIP Model (Vision). If your training is not for "Image-to-Video" (I2V) models, leave it as `None`. If it is I2V, select the appropriate CLIP model, as it helps the model understand the visual information from the input images.

Outputs (3)

NameTypeDescription
dataset_toml_pathSTRING
trigger_out*
status_messageSTRING